CNN and XGBoost for Automatic Segmentation of Stroke Lesions using CT Data

Abstract

The negative impact of stroke on society has led to a concerted effort to improve stroke management and diagnosis. Technological advances have led to the development of deep learning techniques that can be used to accurately detect and characterize brain damage caused by stroke. The ubiquitous growth of artificial intelligence and its medical applications has improved the efficiency of healthcare systems for patients requiring rapid care. Today, chronic diseases such as stroke are the leading cause of death worldwide. A stroke is a type of brain injury. Stroke lesions occur when a group of brain cells dies due to a lack of blood supply. Stroke damage can disrupt brain function, causing a wide range of symptoms such as weakness, disturbance of one or more senses and confusion. It is important to detect and characterize the damaged areas of brain tissue caused by stroke accurately, quickly and effectively in order to save the patient. Stroke diagnosis often relies on expensive imaging techniques such as CT and magnetic resonance imaging (MRI), which are expensive nevertheless CT is more available and accessible in some African hospitals in general and in Senegal in particular. We therefore use a CT dataset to automatically segment stroke lesions. We find that CNN combined with XGBoost can effectively detect, classify and characterize stroke-damaged areas.